Machine learning models to predict red blood cell transfusion in patients undergoing mitral valve surgery

A precise blood transfusion evaluation model for aortic surgery: a single-center retrospective study

Cardiac aortic surgery is an extremely complicated procedure that often requires large volume blood transfusions during the operation. Currently, it is not possible to accurately estimate the intraoperative blood transfusion volume before surgery. Therefore, in this study, to determine the clinically precise usage of blood for intraoperative blood transfusions during aortic surgery, we established a predictive model based on machine learning algorithms. We performed a retrospective analysis on 4,285 patients who received aortic surgery in Beijing Anzhen Hospital between January 2018 and September 2022. Ultimately, 3,654 patients were included in the study, including 2,557 in the training set and 1,097 in the testing set. By utilizing 13 current mainstream models and a large-scale cardiac aortic surgery dataset, we built a novel machine learning model for accurately predicting intraoperative red blood cell transfusion volume. Based on the transfusion-related risk factors that the model identified, we also established the relevant variables that affected the results. The results revealed that decision tree models were the most suitable for predicting the blood transfusion volume during aortic surgery. In particular, the mean absolute error for the best-performing extremely randomized forest model was 1.17 U, while the R2 value was 0.50. Further exploration into intraoperative blood transfusion during aortic surgery identified erythrocytes, estimated operation duration, body weight, sex, red blood cell count, and D-dimer as the six most significant risk factors. These factors were subsequently analyzed for their influence on intraoperative blood transfusion volume in relevant patients, as well as the protective threshold for prediction. The novel intraoperative blood transfusion prediction model for cardiac aorta surgery in this study effectively assists clinicians in accurately calculating blood transfusion volumes and achieving effective utilization of blood resources. Furthermore, we utilize interpretability technology to reveal the influence of critical risk factors on intraoperative blood transfusion volume, which provides an important reference for physicians to provide timely and effective interventions. It also enables personalized and precise intraoperative blood usage.

Development and validation of a machine learning prediction model for perioperative red blood cell transfusions in cardiac surgery

OBJECTIVE

Several machine learning (ML) models have been used in perioperative red blood cell (RBC) transfusion risk for cardiac surgery with limited generalizability and no external validation. Hence, we sought to develop and comprehensively externally validate a ML model in a large dataset to estimate RBC transfusion in cardiac surgery with cardiopulmonary bypass (CPB).

DESIGN

A retrospective analysis of a multicenter clinical trial (NCT03782350).

PATIENTS

The study patients who underwent cardiac surgery with CPB came from four cardiac centers in China and Medical Information Mart for Intensive Cared (MIMIC-IV) dataset.

MEASUREMENTS

Data from Fuwai Hospital were used to develop an individualized prediction model for RBC transfusion. The model was externally validated in the data from three other centers and MIMIC-IV dataset. Twelve models were constructed.

MAIN RESULTS

A total of 11,201 eligible patients were included in the model development (2420 in Fuwai Hospital) and external validation (563 in the other three centers and 8218 in the MIMIC-IV dataset). A significant difference was observed between the Logistic Regression and CatboostClassifier (0.72 Vs. 0.74, P = 0.031) or RandomForestClassifier (0.72 Vs. 0.75 p = 0.012) in the external validation and MIMIV-IV datasets (age ≤ 70:0.63 Vs. 0.71, p < 0.001; age > 70:0.63 Vs. 0.70, 0.63 Vs. 0.71, p < 0.001). The CatboostClassifier and RandomForestClassifier model was comparable in development (0.83 Vs. 0.82, p = 0.419), external (0.74 Vs. 0.75, p = 0.268), and MIMIC-IV datasets (age ≤ 70: 0.71 Vs. 0.71, p = 0.574; age > 70: 0.70 Vs. 0.71, p = 0.981). Of note, they outperformed other ML models with excellent discrimination and calibration. The CatboostClassifier and RandomForestClassifier models achieved higher area under precision-recall curve and lower brier loss score in validation and MIMIC-IV datasets. Additionally, we confirmed that low preoperative hemoglobin, low body mass index, old age, and female sex increased the risk of RBC transfusion.

CONCLUSIONS

In our study, enrolling a broad range of cardiovascular surgeries with CPB and utilizing a restrictive RBC transfusion strategy, robustly validates the generalizability of ML algorithms for predicting RBC transfusion risk. Notably, the CatboostClassifier and RandomForestClassifier exhibit strong external clinical applicability, underscoring their potential for widespread adoption. This study provides compelling evidence supporting the efficacy and practical value of ML-based approaches in enhancing transfusion risk prediction in clinical practice.

Prediction of intraoperative red blood cell transfusion in valve replacement surgery: machine learning algorithm development based on non-anemic cohort

Background

Our study aimed to develop machine learning algorithms capable of predicting red blood cell (RBC) transfusion during valve replacement surgery based on a preoperative dataset of the non-anemic cohort.

Methods

A total of 423 patients who underwent valvular replacement surgery from January 2015 to December 2020 were enrolled. A comprehensive database that incorporated demographic characteristics, clinical conditions, and results of preoperative biochemistry tests was used for establishing the models. A range of machine learning algorithms were employed, including decision tree, random forest, extreme gradient boosting (XGBoost), categorical boosting (CatBoost), support vector classifier and logistic regression (LR). Subsequently, the area under the receiver operating characteristic curve (AUC), accuracy, recall, precision, and F1 score were used to determine the predictive capability of the algorithms. Furthermore, we utilized SHapley Additive exPlanation (SHAP) values to explain the optimal prediction model.

Results

The enrolled patients were randomly divided into training set and testing set according to the 8:2 ratio. There were 16 important features identified by Sequential Backward Selection for model establishment. The top 5 most influential features in the RF importance matrix plot were hematocrit, hemoglobin, ALT, fibrinogen, and ferritin. The optimal prediction model was CatBoost algorithm, exhibiting the highest AUC (0.752, 95% CI: 0.662-0.780), which also got relatively high F1 score (0.695). The CatBoost algorithm also showed superior performance over the LR model with the AUC (0.666, 95% CI: 0.534-0.697). The SHAP summary plot and the SHAP dependence plot were used to visually illustrate the positive or negative effects of the selected features attributed to the CatBoost model.

Conclusions

This study established a series of prediction models to enhance risk assessment of intraoperative RBC transfusion during valve replacement in no-anemic patients. The identified important predictors may provide effective preoperative interventions.

Forecasting demand for blood products: Towards inventory management of a perishable product

Forecasting consumption of blood products can reduce their order frequency by 60% and inventory level by 40%. This also prevents shortage by balancing demand and supply. The study aimed to establish a "Simple Average with Mean Annual Increment" (SAMAI) method of time series forecasting and to compare its results with those of ARIMA, ratio to trend, and simple average to forecast demand of blood products. Monthly demand data of blood component from January 2017 to December 2022 (data set I) was used for creating a forecasting model. To avoid the effect of COVID19 pandemic instead of actual data of year 2020 and 2021, average monthly values of previous three years were used (data set II). The data from January to July 2023 were used as testing data set. To assess the fitness of model MAPE (Mean Absolute Percentage Error) was used. By SAMAI method MAPE were 18.82%, 13.392%, 14.516% and 27.637% respectively for of blood donation, blood issue, RDP issue and FFP issue for data set I. By Simple Average method MAPE were 20.05%, 12.09%, 29.06% and 34.85%, respectably. By Ratio-to-Trend method MAPE were 21.08%, 21.65%, 25.62% and 39.95% respectively. By SARIMA method MAPE were 12.99%, 19.59%, 37.15% and 31.94% respectively. The average MAPE was lower in data set II by all tested method and overall MAPE was lower by SAMAI method. The SAMAI method is simple and easy to perform. It can be used in the forecasting of blood components demand in medical institution without knowledge of advanced statistics.

Using machine learning to predict bleeding after cardiac surgery

OBJECTIVES

The primary objective was to predict bleeding after cardiac surgery with machine learning using the data from the Australia New Zealand Society of Cardiac and Thoracic Surgeons Cardiac Surgery Database, cardiopulmonary bypass perfusion database, intensive care unit database and laboratory results.

METHODS

We obtained surgical, perfusion, intensive care unit and laboratory data from a single Australian tertiary cardiac surgical hospital from February 2015 to March 2022 and included 2000 patients undergoing cardiac surgery. We trained our models to predict either the Papworth definition or Dyke et al.'s universal definition of perioperative bleeding. Our primary outcome was the performance of our machine learning algorithms using sensitivity, specificity, positive and negative predictive values, accuracy, area under receiver operating characteristics curve (AUROC) and area under precision-recall curve (AUPRC).

RESULTS

Of the 2000 patients undergoing cardiac surgery, 13.3% (226/2000) had bleeding using the Papworth definition and 17.2% (343/2000) had moderate to massive bleeding using Dyke et al.'s definition. The best-performing model based on AUPRC was the Ensemble Voting Classifier model for both Papworth (AUPRC 0.310, AUROC 0.738) and Dyke definitions of bleeding (AUPRC 0.452, AUROC 0.797).

CONCLUSIONS

Machine learning can incorporate routinely collected data from various datasets to predict bleeding after cardiac surgery.

Systematic review highlights high risk of bias of clinical prediction models for blood transfusion in patients undergoing elective surgery

BACKGROUND

Blood transfusion can be a lifesaving intervention after perioperative blood loss. Many prediction models have been developed to identify patients most likely to require blood transfusion during elective surgery, but it is unclear whether any are suitable for clinical practice.

STUDY DESIGN AND SETTING

We conducted a systematic review, searching MEDLINE, Embase, PubMed, The Cochrane Library, Transfusion Evidence Library, Scopus, and Web of Science databases for studies reporting the development or validation of a blood transfusion prediction model in elective surgery patients between January 1, 2000 and June 30, 2021. We extracted study characteristics, discrimination performance (c-statistics) of final models, and data, which we used to perform risk of bias assessment using the Prediction model risk of bias assessment tool (PROBAST).

RESULTS

We reviewed 66 studies (72 developed and 48 externally validated models). Pooled c-statistics of externally validated models ranged from 0.67 to 0.78. Most developed and validated models were at high risk of bias due to handling of predictors, validation methods, and too small sample sizes.

CONCLUSION

Most blood transfusion prediction models are at high risk of bias and suffer from poor reporting and methodological quality, which must be addressed before they can be safely used in clinical practice.

How do I/we forecast tomorrow's transfusion? A focus on recipients' profiles

Red blood cell (RBC) transfusion is a life-saving medical intervention and has an essential role in the management of surgical patients. However, blood donations and supply levels are decreasing, therefore there is an unmet need for the accurate prediction of the transfusion probability for surgical patients. Multiple methods have been established to predict the need for RBC transfusion. Maximum surgical blood order schedules are widely used in the clinical setting. However, these lists are not designed to accurately predict RBC utilization for an individual case as factors such as preoperative haemoglobin level, total body blood volume, comedications are not considered. Artificial intelligence and related technologies based on machine learning modelling are valuable alternatives to predict transfusion probability taking into account patient individual risk factors including among others comorbidities, laboratory parameters, use of oral anticoagulation, ASA score, surgeon's ID or applied blood saving measures. Overall, forecasting the need for a RBC transfusion can facilitate personalized medicine, quality assurance, decrease blood wastage, decrease costs, and increase patient safety. Furthermore, transfusion prediction models could facilitate blood management strategies before surgery.

Multivariable statistical models to predict red cell transfusion in elective surgery

BACKGROUND

Predicting red cell transfusion may assist in identifying those most likely to benefit from patient blood management strategies. Our objective was to identify a simple statistical model to predict transfusion in elective surgery from routinely available data.

MATERIALS AND METHODS

Our final multicentre cohort consisted of 42,546 patients and contained the following potential predictors of red cell transfusion known prior to admission: patient age, sex, pre-admission hemoglobin, surgical procedure, and comorbidities. Missing data were handled by multiple imputation methods. The outcome measure of interest was administration of a red cell transfusion. We used multivariable logistic regression models to predict transfusion, and evaluated the performance by applying a 10-fold cross-validation. Model accuracy was assessed by comparing the area under the receiver operating characteristics curve. After applying an optimal probability cut-off we measured model accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

RESULTS

7.0% (n=2,993) of the study population received a red cell transfusion. Our most simple model predicted red cell transfusion based on admission hemoglobin and surgical procedure with a multiply imputed estimated area under the curve of 0.862 (0.856, 0.864). The estimated accuracy, sensitivity, specificity, positive predictive, and negative predictive values at the probability cut-off of 0.4 were 0.934, 0.257, 0.986, 0.573, and 0.946 respectively.

DISCUSSION

A small number of variables available prior to admission can predict red cell transfusion with very good accuracy. Our model can be used to flag high-risk patients most likely to benefit from pre-operative patient blood management measures.

Prediction of Red Blood Cell Demand for Pediatric Patients Using a Time-Series Model: A Single-Center Study in China

Background

Red blood cells (RBCs) are an essential factor to consider for modern medicine, but planning the future collection of RBCs and supply efforts for coping with fluctuating demands is still a major challenge.

Objectives

This study aimed to explore the feasibility of the time-series model in predicting the clinical demand of RBCs for pediatric patients each month.

Methods

Our study collected clinical RBC transfusion data from years 2014 to 2019 in the National Center for Children's Health (Beijing) in China, with the goal of constructing a time-series, autoregressive integrated moving average (ARIMA) model by fitting the monthly usage of RBCs from 2014 to 2018. Furthermore, the optimal model was used to forecast the monthly usage of RBCs in 2019, and we subsequently compared the data with actual values to verify the validity of the model.

Results

The seasonal multiplicative model SARIMA (0, 1, 1) (1, 1, 0)₁₂ (normalized BIC = 8.740, R² = 0.730) was the best prediction model and could better fit and predict the monthly usage of RBCs for pediatric patients in this medical center in 2019. The model residual sequence was white noise (Ljung-Box Q₍₁₈₎ = 15.127, P > 0.05), and its autocorrelation function (ACF) and partial autocorrelation function (PACF) coefficients also fell within the 95% confidence intervals (CIs). The parameter test results were statistically significant (P < 0.05). 91.67% of the actual values were within the 95% CIs of the forecasted values of the model, and the average relative error of the forecasted and actual values was 6.44%, within 10%.

Conclusions

The SARIMA model can simulate the changing trend in monthly usage of RBCs of pediatric patients in a time-series aspect, which represents a short-term prediction model with high accuracy. The continuously revised SARIMA model may better serve the clinical environments and aid with planning for RBC demand. A clinical study including more data on blood use should be conducted in the future to confirm these results.

Identification and Validation of the Pyroptosis-Related Molecular Subtypes of Lung Adenocarcinoma by Bioinformatics and Machine Learning

Lung cancer remains the leading cause of cancer death globally, with lung adenocarcinoma (LUAD) being its most prevalent subtype. Due to the heterogeneity of LUAD, patients given the same treatment regimen may have different responses and clinical outcomes. Therefore, identifying new subtypes of LUAD is important for predicting prognosis and providing personalized treatment for patients. Pyroptosis-related genes play an essential role in anticancer, but there is limited research investigating pyroptosis in LUAD. In this study, 33 pyroptosis gene expression profiles and clinical information were collected from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases. By bioinformatics and machine learning analyses, we identified novel subtypes of LUAD based on 10 pyroptosis-related genes and further validated them in the GEO dataset, with machine learning models performing up to an AUC of 1 for classifying in GEO. A web-based tool was established for clinicians to use our clustering model (http://www.aimedicallab.com/tool/aiml-subphe-luad.html). LUAD patients were clustered into 3 subtypes (A, B, and C), and survival analysis showed that B had the best survival outcome and C had the worst survival outcome. The relationships between pyroptosis gene expression and clinical characteristics were further analyzed in the three molecular subtypes. Immune profiling revealed significant differences in immune cell infiltration among the three molecular subtypes. GO enrichment and KEGG pathway analyses were performed based on the differential genes of the three subtypes, indicating that differentially expressed genes (DEGs) were involved in multiple cellular and biological functions, including RNA catabolic process, mRNA catabolic process, and pathways of neurodegeneration-multiple diseases. Finally, we developed an 8-gene prognostic model that accurately predicted 1-, 3-, and 5-year overall survival. In conclusion, pyroptosis-related genes may play a critical role in LUAD, and provide new insights into the underlying mechanisms of LUAD.